TEGKT: tendency-enhanced evolution graph KAN transformer for information popularity prediction
Abstract According to historical retweet relationships that reveal public attention, information popularity prediction aims to forecast the incremental size of the given information cascades. Existing work independently models user dynamic preference with discrete cascade snapshots technology, they...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00170-8 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract According to historical retweet relationships that reveal public attention, information popularity prediction aims to forecast the incremental size of the given information cascades. Existing work independently models user dynamic preference with discrete cascade snapshots technology, they ignore the global structure of information cascades and inefficient tendency semantic representation, leading to suboptimal performance. To overcome the those issues, we introduce a novel Tendency-Enhanced Evolution Graph KAN Transformer framework ( TEGKT ), which is specifically tailored for information popularity prediction. To enhance the ability to express tendency semantics, we construct a tendency encoding learning module, which can effectively exhume the potential high-level dependency relationship among tendency semantics. To capture the global structure of cascade snapshots during the observation period, we design the evolution Graph KAN Transformer architecture to improve the expressive ability of information cascade representation, and its weighted parameter is optimized by gated recurrent units (GRU). Bi-directional gate recurrent units (Bi-GRU) are used to explore the dynamic evolution between cascade snapshots. Extensive experiments conducted on three public datasets show that the proposed model significantly outperforms the advanced methods, which achieves by 3.34% and 4.09% on the Weibo 0.5h dataset in terms of MSLE and MAPE evaluation metrics, respectively, validating its effectiveness. The research provides a better understanding of the laws of information diffusion. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |